I am a Ph.D. candidate in Data Science at Seoul National University, advised by Prof. Taesup Kim in the Learning and Adaptation Algorithms Lab.

I study post-training of multimodal foundation models: how they acquire new capabilities without losing existing ones, and whether their reasoning stays grounded in visual evidence. Most recently, I completed a five-month research internship at LG AI Research’s EXAONE Lab, where I studied reinforcement-learning post-training for general-purpose vision-language models.

Before entering research, I worked as a journalist at The Chosun Ilbo. In both journalism and research, I have been happiest in conversations that help us understand one another and explore the world and intelligence, whether human or artificial.

I am looking for research internships for Summer 2027.

Research

My work studies tensions in foundation-model training: preserving old knowledge while acquiring new capabilities, following language priors while remaining grounded in visual evidence, and allocating learning signal across competing capabilities.

Prior knowledge New knowledge

Continual learning without replay or parameter growth (ICLR 2026)

Language priors Visual evidence

Training-free attention steering against hallucination (CVPR 2026 Findings)

Reasoning-heavy tasks Perception-heavy tasks

Allocating learning signal across competing capabilities in multi-task RL (NeurIPS 2026 Workshops)

News

  1. Our work on hidden curricula in vision-language RL was accepted at two NeurIPS 2026 workshops, VLM4RWD (Poster) and Pre-to-Post.
  2. Finished my research internship at LG AI Research's EXAONE Lab.
  3. NuSA-CL was accepted at ICLR 2026, and ACG was accepted at CVPR 2026 Findings.

Selected Publications

View all

NeurIPS 2026 Workshops

When Learning Signals Persist: Hidden Curricula in Vision-Language Reinforcement Learning

Yujin Jo, Hyunjik Jo, Yireun Kim, Jinsik Lee, Taesup Kim

Online filtering in group-relative RL turns a fixed data mixture into a hidden curriculum; the queries that remain gradient-eligible longest are not necessarily the most useful to train on.

Camera-ready version coming soon

ICLR 2026

Memory-Free Continual Learning with Null Space Adaptation for Zero-Shot Vision-Language Models

Yujin Jo, Taesup Kim

Treats low-energy spectral directions as writable capacity, enabling continual adaptation without replay, a teacher model, task-specific modules, or parameter growth.

Paper Code

CVPR 2026 Findings

Attention-space Contrastive Guidance for Efficient Hallucination Mitigation in LVLMs

Yujin Jo, Sangyoon Bae, Taesup Kim

A single-pass, training-free method that contrasts vision-conditioned and text-only attention paths to keep answers grounded in visual evidence.

Paper Code

Experience

LG AI Research, EXAONE Lab

Research Intern

RL post-training for general-purpose vision-language models.

Learning and Adaptation Algorithms Lab, SNU

Graduate Researcher

Reliable and efficient adaptation and post-training of vision-language models.

The Chosun Ilbo

Journalist

Reported primarily on social affairs, with additional coverage of technology and science.

Education

Seoul National University

Ph.D. Candidate in Data Science

Advisor: Prof. Taesup Kim · GPA: 4.25/4.3

Seoul National University

M.S. in Data Science

Thesis: Group-wise Prompt Learning for Vision-Language Models · GPA: 4.06/4.3

Korea Advanced Institute of Science and Technology (KAIST)

B.S. in Electrical Engineering

Service & Recognition

  • Reviewer for NeurIPS 2026 and ICLR 2027
  • Lecture & Research Scholarship, Seoul National University (2023, 2024)
  • KAIST Presidential Fellowship (2012 - 2016)
  • Talent Medal of Korea, awarded by the President of the Republic of Korea (2011)